Hermes or Grok as the agent, Claude Code or ChatGPT as the workhorse — and the context underneath that turns them into digital teammates
Most companies have the tools and still get generic output, because the agent knows nothing about the company. We build the context layer: persona files for the CEO, CTO, engineers, sales, HR and hiring; connected inboxes; a read-only, sanitised view of your user database; recorded and indexed transcripts; memory that extends instead of resets; and plugins that keep it current. Then we wire it to a human-in-the-loop draft pipeline — the one two people use to run support for 40,000+ users at Pocket Clear. Fixed price, from $1,000.
40,000+
Users supported by a two-person team
114
Requests tallied from email + store reviews
40
Shipped from the tally, each with a 'you asked, it's live' email
28
Reply drafts written in two weeks — every one read before sending
Every team we talk to has already tried it: a Hermes agent on the inbox, a Grok bot on X, Claude Code or ChatGPT in the repo or on the docs. The output is generic, so it gets ignored. The agent does not know that the feature the customer is asking about shipped in 2.09 last week, that this account is on the annual plan, that the CTO decided against that integration in March, or that the CEO never uses the word 'synergy'. Nobody wrote it down anywhere an agent can read. That is the context problem, and it is why most AI employees are still demos rather than digital teammates.
The context layer is what we build. It is not one thing; it is a set of sources the agent can read and a set of rules for keeping them current. Persona files — one per role, so the CEO's digital teammate and the HR teammate know different things and speak differently. Connected inboxes, read-only. A read-only, sanitised view of your user database, so a draft can check account state without ever seeing raw PII. Meeting and call transcripts recorded, indexed and summarised so decisions stop living in people's heads. Memory that extends — a tally, a decision log, a contact log — instead of resetting every session. Plugins and skills that pull the current state of your repo, your CRM, your help centre. And abstraction rules: what to summarise, what to keep verbatim, what never to store.
On top of that layer, the pipeline. Hermes or Grok reads everything new — support email, store reviews, the channels your customers use — classifies it, extracts the ask in one line, appends a row to a tally, and flags what needs a human now. Claude Code or ChatGPT, with the context layer loaded, drafts every reply into a dated drafts/ folder and, for startups, implements what the tally says to build. Nothing customer-facing is sent by a machine. At Pocket Clear about 80% of drafts go out as written, 20% get edited, and the tally replaced the roadmap — see the write-up (/insights/ai-employee-for-customer-support) and the context-layer guide (/insights/ai-employee-context-layer).
Once the context layer exists, other jobs become cheap. The same agent that reads your inbox can watch X, Reddit, Instagram and TikTok for the topics your customers are already talking about and hand you a weekly list of what to post, with drafts in your voice — because it knows your voice (/insights/ai-content-radar-x-reddit-instagram-tiktok). Sales outreach tuned on your own closed-won threads and the replies you actually edited, hiring, payroll questions, investor updates: each is a persona file and a source away (/insights/ai-employee-context-by-persona).
We only install what we run ourselves — Hermes, Grok bots, Claude Code, ChatGPT — on our own products, and we hand over a runbook so you are not dependent on us. Security is a line item in every package, not an upsell: the agent runs on hardware you control, credentials never sit in a config file, database access is read-only and sanitised, and nothing auto-sends (/insights/securing-an-ai-employee-on-your-inbox).
from $1,000one-time · $1K–$3K by personas and flows
Best for: Freelancers, solo operators, and small businesses — one persona, one inbox
One persona's context layer, one agent, one flow. Live in 2–3 days. The morning read-tally-flag routine with drafts, hardened, with a runbook so you own it.
from $10,000one-time · $10K–$15K by personas, sources, and flows
Best for: Startups with a product, a team, a user database, and a repo
The company context layer: personas for CEO, CTO, engineers, sales, HR/payroll and hiring; inboxes, sanitised database, transcripts and repo as sources; the full read → tally → Claude Code build-and-draft pipeline; and the content radar. Live in 2 weeks.
Running agent (Hermes or Grok) on hardware you control, with Claude Code or ChatGPT connected and a documented restart and update procedure
Persona files per role, memory files, and routines checked into a workspace you own
Connected inboxes (read-only) and, for Growth, a sanitised read-only database view with the masking rules documented
Transcript pipeline: how recordings are captured, indexed, summarised, and where the decision log lives
Classification rules, tally format, and urgent-flag rules tuned to your product's vocabulary
Security runbook: credential storage, network exposure, data access, log retention, what the agent can and cannot read
Human-in-the-loop map: exactly which steps produce a draft versus an action
Walkthrough recording, written handover, and a tuning window to fix misclassifications and merge duplicate rows
We run this on our own product — Pocket Clear, 40,000+ users, 2 people — not a demo account
We sell the context layer, not a tool: Hermes, Grok bots, Claude Code, ChatGPT — whichever you already use, made to know your company
Per-persona from day one: a CEO's digital teammate and an HR teammate should not know the same things or sound the same
Data access is read-only and sanitised by design, so connecting the database is safe rather than scary
The tally is the product. A vote count told us our users cared 3% about the thing we led every pitch with
Fixed price, fixed scope, one-time. No per-seat subscription, no retainer. A founder sets it up and hands it over
Day 0
A 30-minute call to see what the agent would need to know and where that knowledge lives today — inboxes, database, meetings, repo, people's heads.
Days 1–10
Persona and memory files written, sources connected read-only, database view sanitised, transcripts indexed, security checklist applied before the first real email is read.
Days 11–30
You review real drafts for two to four weeks while we fix misclassifications, extend memory, and write the runbook.
Days 0–3
Starter goes live. Growth has the reading half and the first persona running.
Days 4–14
Growth: remaining personas, database view, transcripts, repo.
Days 15–30
Tuning, content radar, handover.
Proof it works
Hermes reads every support email and App Store / Google Play review into a research tally; Claude Code, with the product's context layer loaded, turns the tally into shipped features and reply drafts that know what shipped last week and what the account looks like. Nothing is auto-sent. The tally replaced the roadmap — and revealed that 'privacy-first', the pitch we led with for a year, was 3% of what users actually talked about.
40,000+
Users, 195 countries
114 → 40
Requests tallied → shipped
11%
Reviews that mention 'I emailed and it got built'
2
Team size, still
“We didn't build a support bot. We built a system that reads everything, counts it, and hands the count to the thing that can build. The human is still in the loop at the only point that matters: the send button.”
Aravind Srinivas
Founder, Pocket Clear & HyperNest Labs
What is the context layer, and what makes an AI employee a digital teammate?
Everything an agent needs to know about your company that is not in the model: who it is speaking as (persona files per role), what has happened (indexed transcripts and a decision log), what is true right now (read-only inboxes, a sanitised database view, the repo, the help centre), what it has already done (a tally and contact log that extend rather than reset), and what it must never store. Without it, Claude Code, ChatGPT, Hermes and Grok all produce generic output. With it, drafts arrive knowing that the feature shipped in 2.09 and the customer is on the annual plan — that is the difference between an AI employee you supervise and a digital teammate you trust.
Which tools do you set up — Hermes, Grok, Claude Code, or ChatGPT?
Whichever you already use, or the pair that fits. Hermes or a Grok bot as the agent that reads, classifies and tallies; Claude Code or ChatGPT as the workhorse that drafts and, in a repo, builds. The context layer is the same underneath either pair. We only install tools we run in production on our own products.
How does the agent use our user database safely?
Through a read-only view we build for it, with PII masked: names and emails hashed or truncated, payment details excluded, free-text fields stripped. The agent can answer 'is this account on the annual plan and when did it last sync' without ever seeing the raw row, and it physically cannot write. The masking rules are documented in the runbook so your team and, if relevant, your compliance counsel can review them.
What does 'per persona' mean in practice?
A separate persona file and source set per role. The CEO's agent knows the metrics, the investor list, the narrative and the voice, and drafts updates. The CTO's knows the architecture decisions and the repo, and drafts RFC responses. The HR/payroll agent knows policies and pay cycles and answers employee questions from them. The sales teammate knows the closed-won threads, the objections, and the contact log, and drafts outreach tuned on what actually got replies. The hiring teammate knows the scorecards and the pipeline. They share the company decision log but not each other's data. The guide is at /insights/ai-employee-context-by-persona.
What does it cost, and is there a subscription?
Starter is from $1,000 one-time ($1K–$3K by personas and flows): one persona, one inbox, one agent, one flow, live in 2–3 days. Growth is from $10,000 one-time ($10K–$15K): the company context layer for up to six personas, sanitised database view, transcripts, repo integration, the full pipeline, sales outreach tuning and the content radar, live in 2 weeks. No subscription, no per-seat fee. You pay your model and API costs directly.
Is it secure? What happens when it gets something wrong?
The agent runs on hardware you control with no inbound exposure; keys live in the OS keychain or environment, never in a config file; mail scopes are read and draft only; database access is read-only and sanitised; nothing customer-facing is sent by a machine. It will still misclassify — roughly once a week in our own setup — and drafts need edits about 20% of the time. The tuning window exists for that, and anything emotional, lost-data or legal is routed to a person and never gets a draft.
Pick a time below — we'll scope your situation and map an engagement.
14-day risk-free trial · usage-based, no retainers · you talk to the founder, not a sales rep
We'll audit your architecture, map out an engagement, and plug in team members within days.
Book a 30-min context audit